Avataar AI launches Varya, a low‑cost video model for India’s market

Avataar AI, one of twelve startups funded by India’s AI Mission, has unveiled Varya, a video‑generation model designed to understand Indian cultural contexts and operate at a fraction of the cost of existing solutions [1]. The model, derived from Alibaba’s Wan 2.2 through distillation, promises faster output and cheaper pricing for developers and enterprises.

Key takeaways

• Varya runs in four steps versus Wan 2.2’s 50, delivering video 10 times faster [1]
• On an NVIDIA H200 GPU it creates a 5‑second 720p clip in 45 seconds, compared with 1,230 seconds for the original model [1]
• Avataar plans to charge ₹0.48 ($0.005) per second of generated video, roughly 20 times cheaper than competing services [1]
• The model will be released as an open‑weight asset on India’s AI Kosh portal, with training data available for self‑hosting [1]
• Avataar’s approach reflects a broader Indian strategy to focus on task‑specific AI applications rather than large foundation models [2]

Distilling Speed and Cost for the Indian Market

Avataar AI leveraged the publicly available Wan 2.2 video generation model from Alibaba and applied a distillation technique to compress its capabilities into a leaner “student” model optimized for e‑commerce video tools [1]. This process reduced the number of inference steps from 50 to just four, cutting generation time by an order of magnitude. Using a high‑end NVIDIA H200 GPU, Varya can produce a 5‑second, 720p video in 45 seconds, whereas Wan 2.2 requires over 20 minutes for the same output [1]. The company markets the service at ₹0.48 per second of video, a price point it claims is about 20 times lower than rivals such as Veo, Kling, Luma, and Runway, which typically charge $0.10 or more per second [1].

Building Cultural Awareness and Open Access

Beyond speed and price, Avataar AI emphasizes cultural relevance. The firm says it curated a dataset that teaches Varya to recognize Indian festivals, food, clothing, architecture, and other local nuances, addressing a known shortfall in many image and video generators that often produce stereotyped outputs [1][2]. Varya will be published as an open‑weight model on the government‑run AI Kosh portal, together with its training data, enabling developers to self‑host, modify, or integrate the model into their own products [1][2]. Avataar also plans to offer the model to enterprise customers and is open to partnerships with video‑tool providers such as Higgsfield and Adobe Firefly [1][2].

Why it matters

Varya’s launch underscores India’s strategic shift toward task‑specific AI that can scale affordably for a video‑first market. Industry observers note that India’s AI ambitions rely more on building applications and a robust developer ecosystem than on competing with global foundation‑model giants, a stance driven by limited compute resources and data quality challenges [2]. The India AI Mission, a $1.2 billion government program, supports startups like Avataar with subsidized GPU compute in exchange for publicly releasing models, aiming to accelerate home‑grown AI innovation and attract $200 billion in AI investment by 2028 [1][2]. As Varya becomes publicly available, its lower cost and cultural awareness could enable broader adoption of AI‑generated video across education, small‑business, and public‑service sectors, potentially setting a template for future Indian AI initiatives.